Downloaded artifacts from terminal-bench CI runs should not be committed.

In-place workspaces (identified by projectPath === workspaceName) are direct
workspace directories used by CLI/benchmark sessions, not git worktrees. Attempting
to run 'git worktree remove' on them fails or attempts to remove the main checkout.

This fix detects the in-place sentinel pattern and skips git worktree operations,
allowing session cleanup without destructive filesystem operations.

Resolves Codex review comment in PR #472.

When the current branch has no upstream, automatically run git push -u
to set it instead of failing. This makes the script more user-friendly
for new branches.

- Run full benchmark suite (~80 tasks) every night at midnight UTC
- Concurrency=4 is appropriate for full suite (60-90 min estimated)
- Timeout=180 min (3 hours) provides safety margin
- Use default fallbacks for scheduled runs (no inputs)
- Add unique artifact names with run_id to avoid conflicts
- Set 30-day retention for nightly benchmark artifacts

terminal-bench-core==0.1.1 contains ~80 tasks, which is the complete
stable benchmark suite. The -head version is bleeding-edge dev.

Use matrix strategy to run both models every night:
- anthropic:claude-sonnet-4-5 (high thinking)
- openai:gpt-5-codex (high thinking)

Matrix only applies to scheduled runs (cron), not manual workflow_dispatch.
Artifacts are named uniquely per model to avoid conflicts.

This enables direct comparison of model performance on the full 80-task suite.

Cleaner architecture:
- terminal-bench.yml: Reusable workflow (workflow_call + workflow_dispatch)
- nightly-terminal-bench.yml: Scheduled runner with matrix strategy

Benefits:
- Main workflow stays simple for manual use
- Nightly schedule logic isolated in dedicated file
- Easy to add more models to nightly runs
- Manual workflow_dispatch supports model/thinking overrides

Nightly runs both models at midnight UTC:
- anthropic:claude-sonnet-4-5 (high thinking)
- openai:gpt-5-codex (high thinking)